With the rapid development of modern Internet of Things (IoT) technology, thehandwritten signature verification system becomes a typical Human-ComputerInteraction (HCI) application, which is often applied to many authorization documents with legal uses. To build a high-performance verification system, featureextraction is one of the most crucial steps. However, the generalization abilityof traditional deep learning-based feature extractors is not always satisfactorysince most feature extractors are only suitable for source signature datasets.In this paper, to improve the generalization ability of existing deep learning-based feature extractors for the offline signature verification task, we proposea novel multi-view learning-based framework, named Deep Canonically Corre-lated Denoising Autoencoders (DCCDAE). Specifically, the DCCDAE generatesthe joint features as the final features based on original deep learning-basedfeatures and another noisy view of them by optimizing the Canonically Cor-related Analysis (CCA) objective and minimizing the reconstruction error oforiginal features. Extensive experiments and discussions on four publicly avail-able datasets, GPDS, CEDAR, MCYT-75, and PUC-PR demonstrate that theproposed DCCDAE can improve the generalization ability of the original deeplearning-based feature extractor and achieve the state-of-the-art performancecompared with other offline signature verification systems. The code is availableat https://github.com/star0511/DCCDAE